World Models, JEPA, and What They Mean for Vector Databases

We are in the midst of a golden AI summer. The public release of ChatGPT in November 2022 accelerated a wave of adoption that brought Large Language Models (LLMs) into mainstream use across industries. LLMs took AI beyond specialist teams, sparking widespread use and a new ecosystem of supporting infrastructure.

LLMs are still finding new applications, from agents to automation. But in parallel, researchers are exploring where AI can go beyond language. Scaling language models made them powerful, but scaling alone is facing questions about diminishing returns, where costs rise faster than capabilities improve.

As AI moves beyond language, the infrastructure that supports it will need to evolve. Vector databases, which play an important role in how organisations get value from LLMs today, will be directly affected. This article examines one of the most significant developments in this shift: world models, and the JEPA architecture.

While still in the research phase, world models are already revealing what future infrastructure demands might look like. At this early stage, the implication for vector databases seems positive: if these models externalise memory as embeddings, vector workloads will grow.

How retrieval works today

Large language models work by predicting the next token in a sequence. Everything they know is encoded in the parameters learned during training. This works remarkably well, but it has known limitations: training data goes stale, proprietary information is absent, and there is no way to audit where a specific answer came from.

Retrieval-Augmented Generation (RAG) exists to fill these gaps. A vector database stores information as numerical representations called embeddings. When a user asks a question, the system finds the most relevant stored information and provides it to the model alongside the question. This gives the model access to current, specific, and citable information that its parameters alone cannot provide.

World models and JEPA: learning to predict representations, not tokens

World models have existed as a concept in AI for decades. JEPA (Joint Embedding Predictive Architecture), proposed by Yann LeCun in 2022, is one of the most prominent current approaches to building them. It converts visual input into numerical representations, much like LLMs do with text. But where LLMs predict the next word, JEPA predicts abstract representations of what comes next.

The practical result is a model that captures dynamics, object interactions, and physical relationships rather than surface-level patterns (i.e., how they behave rather than how they look). The Meta V-JEPA 2 model, trained on a dataset of over a million hours of video, learned how objects move, interact, and persist (even tracking them when they pass behind other objects). That physical understanding then transferred to a completely different domain: with just 62 hours of robot video, the model could control robots in environments it had never seen.

JEPA and the scale of vector data

Current production embedding models commonly produce a single vector per chunk of content, typically ranging from 384 to 3,072 dimensions depending on the model. A text document might be split into dozens of chunks, each getting one vector. This is already a meaningful amount of data at scale, but the volumes involved are manageable.

By contrast, a single V-JEPA 2 video sequence can generate up to 8,192 internal embedding vectors in a standard configuration, with counts varying by resolution and sequence length. Most of these are working memory, generated during prediction and discarded.

If these systems externalise long-term memory in embedding form, through episodic recall, agent experience, or multimodal retrieval, the volume of stored vector data could grow significantly beyond what text-based RAG produces. These embeddings would also be structurally different: temporally sequenced and capturing how objects behave in the physical world.

For vector databases, this means retrieval is no longer just about finding the most similar chunk of text. It could involve searching across sequences of states, retrieving experience by time and context, and doing so fast enough to inform a model that is actively making decisions.

Industry analysts project the vector database market will more than triple to USD 8.94 billion by 2030, based on current use cases. World model workloads are not yet explicitly called out in these projections, which suggests the actual requirements could be higher.

Why world models still need external memory

It is important to distinguish between two types of memory in these systems. Short-term planning, where a world model simulates possible futures to choose an action, typically happens in temporary memory within the model itself. It does not strictly require an external database.

But long-term memory is a different problem. World models learn how things behave, not what specific things exist. As these systems move toward operating over longer periods, they will need to recall past experiences to adapt, not just simulate the immediate future. For grounding in facts, for recency, for episodic memory of past states and outcomes, external retrieval is likely to remain necessary. Text-based RAG is not going away. It gets joined by new workloads driven by how world models process and store experience.

Where this leaves vector databases

While the research is still unfolding, the direction is relatively clear. As world models mature, the volume and variety of vector data will grow significantly. These are workloads that did not exist when vector databases were first adopted for RAG, and they will require new thinking about how vector data is stored and retrieved.

Vector databases already serve a wide range of AI workloads. World models could expand their role even further. The models are learning to understand the physical world. The infrastructure will need to keep up.

We will keep a close eye on how this research develops and what it means for storing and retrieving vector data at scale, with a focus on what can be applied in practice.